Key Engineering Materials Vol. 1050

Abstract: Hybrid twins and residual-learning strategies are increasingly used to reconcile the broad coverage of physics-based simulations with the fidelity of production measurements. In industrial stamping, however, truly matched simulation-experiment cases are scarce, while large-scale monitoring data are clustered around a narrow operating window. This work proposes an industrially practical hybrid metamodel in which the discrepancy (ignorance) model is trained exclusively from approximate residuals computed at in-domain production inputs, considering a surrogate of a thermal enabled AutoForm Sigma Design of Experiments (DoE). To prevent uncontrolled extrapolation, learning and evaluation are restricted to an explicitly validated domain defined through multivariate kNN support in standardized space derived from the DoE cloud. The residual model is selected through five-fold cross-validation on 10,446 in-domain production samples and then retrained on the full approximate-residual set. A small set of five matched cases is kept as an external check based on true residuals. The resulting hybrid predictor enables the generation of synthetic, experiment-informed corrected data while retaining the DoE coverage required for downstream modelling tasks.

Abstract: Accurate simulation of material forming requires managing severe mesh distortions to preserve the geometry of the workpiece. Classical Lagrangian descriptions often become computationally expensive under such conditions. The Arbitrary Lagrangian-Eulerian (ALE) method offers a robust alternative by combining the strengths of Lagrangian and Eulerian descriptions, thereby improving computational efficiency and numerical stability.However, a major challenge remains in accurately handling free surfaces to maintain the geometric fidelity of the workpiece.This work introduces a new ALE-based approach to address this limitation. An analytical case as well as a Friction Stir Welding (FSW) case will be presented to demonstrate its effectiveness.

Abstract: A novel solution for monitoring the infusion process and providing decision support to operators involved in the manufacturing of large, unique or near-unique parts is presented. Based on a scientific approach referred to as the 5D methodology (D for dimensions), the proposed solution consists of a process digital twin built upon a metamodel that is fed in real time by signals from sensors embedded in the process, enabling the anticipation of defects such as dry spots.

Abstract: Draw-in distance is a key index for evaluating the quality of sheet metal stamping. Its accurate prediction is therefore required for tool design and process control. Traditional finite element (FE) simulations, while accurate, are computationally intensive and time-consuming for iterative design optimization. In this study, a graph neural network (GNN) method is proposed to predict draw-in during sheet metal forming. A dataset was built from FE simulations with different process settings, including blank holder force and draw bead force. The GNN model uses node coordinates and edge features to describe the spatial relations in the sheet. A multi-level loss function was applied. The coordinate error and edge distance error were included. In this way, the shape of the sheet is better preserved. The trained GNN can be used as a fast model for draw-in prediction. It can also be used for inverse analysis, where the process parameters are found from a given draw-in result. This provides an efficient tool for sheet metal forming design and optimization.

Abstract: In this study, we develop a Bayesian data assimilation framework that combines a mean-field model of static recrystallization (MiReX) with a Sequential Importance Resampling (SIR) particle filter to estimate key material parameters from controlled synthetic experiments. MiReX, originally developed as a microstructurally based extension of Johnson–Mehl–Avrami–Kolmogorov kinetics, is used as a forward model in which the uncertain quantities include the grain-boundary mobility parameters (prefactor and activation energy), a stored-energy coefficient, an Avrami-type exponent, and an interface length scale. Synthetic recrystallized-fraction measurements are generated at two isothermal holding temperatures using a reference parameter set and are perturbed with Gaussian noise to mimic experimental uncertainty. Starting from broad uniform prior ranges, the particle filter propagates an ensemble of MiReX trajectories in time, updates particle weights using a Gaussian likelihood, and applies systematic resampling combined with Liu–West kernel regularization to reduce particle degeneracy while preserving posterior variance. The posterior obtained after assimilating the first temperature dataset is used as the prior for the second dataset, enabling sequential multi-temperature calibration. The synthetic experiments show that the framework recovers the reference parameters within credible intervals and provides tight uncertainty bounds on the predicted recrystallization kinetics. These results demonstrate that combining a physically based mean-field recrystallization model with sequential Monte Carlo methods provides a robust route for probabilistic parameter estimation and uncertainty quantification in microstructure evolution models.

Abstract: This work investigates the use of symbolic regression (SR) to address the trade-off between predictive accuracy and computational efficiency in modeling physical phenomena by constructing compact, closed-form expressions directly from data. In this study, SR is applied to develop fast and accurate models for predicting lateral spread in the hot rolling of steel slabs. The SR models are trained on high-fidelity finite element simulation data and evaluated against established analytical models. Model selection is guided by a parsimony-based optimization strategy that balances predictive accuracy and expression complexity. The results show that the SR-derived formulations achieve lower prediction errors with reduced complexity compared to traditional analytical models. Moreover, SR maintains strong predictive performance even when trained on limited datasets, demonstrating its robustness. Overall, the findings of this work highlight the suitability of symbolic regression for computationally efficient and accurate modeling of complex physical phenomena.

Abstract: Metal-forming is a manufacturing process that involves non-linear elastoplastic deformations to shape a blank into a complex geometry. These processes are governed by numerous parameters, with significant influence on the final product. To analyse the effects of such parameters, large-scale finite element (FE) simulations are often conducted. However, these models are computationally expensive and often unsuitable for real-time analysis. To overcome these limitations, surrogate models have emerged as powerful alternatives. In this study, we propose a physics-informed recurrent neural network framework (PIRNN) to evaluate displacements and strain tensor components. In particular, the latter are not a network output but are obtained through the application of kinematic relations. Given the initial configuration as input and the final configuration as output, it is possible to evaluate the deformation gradient. The impenetrability condition is then injected into the loss to improve the estimation of displacements and strain tensors. The PIRNN model, referred to as a kinematics-informed recurrent neural network, is trained on data generated from FE simulations of a deep-drawing process. The accuracy of the model is evaluated on a test dataset (design points that the model does not see during training) using different error measures. The results show that the proposed KI-RNN model can fairly reproduce the FE simulation results fairly well.

Abstract: Predicting the microstructural state during manufacturing is critical, as it directly governs the material's final mechanical properties. Accurate prediction of microstructure evolution in multi-stage industrial hot deformation processes, such as rolling, is limited by the lack of experimental data at intermediate stages, where direct measurement is impractical. To address this, an integrated methodology combining finite element (FE) simulation in QForm UK® software, physical simulation using the Thermo-Mechanical Treatment Simulator (TMTS), and artificial intelligence (AI) is proposed and investigated. The methodology is demonstrated for the 11-pass hot rolling of a 41Cr4 steel bar. Thermomechanical loading histories from an FE model of the industrial process were used to design and simulate a targeted TMTS experiment, generating a synthetic dataset via an analytical JMAK model that combines multiple recrystallisation mechanisms. This data was used to train a recurrent neural network (RNN) with an augmented physics-informed Long Short-Term Memory (LSTM) cell to predict the totally recrystallised fraction (RX) solely from loading history data. The AI model achieved high accuracy when validated within the TMTS simulation domain, successfully capturing different recrystallisation regimes. Implementation within commercial FE software enabled direct prediction in the rolling process simulation, yielding promising predictive capability, particularly in regions with thermal histories similar to the training data, highlighting the critical importance of training data diversity. This work establishes a proof of concept for a novel calibration methodology, where targeted physical simulation bridges the gap between industrial process complexity and data-driven AI model development, offering a practical solution for modelling scenarios where traditional experimental calibration is infeasible.

Abstract: A new efficient numerical solver inspired by front-tracking concepts is implemented within the DIGIMU® framework to accelerate full-field simulations of microstructural evolution. The solver is applied to AISI 304L stainless steel and compared with the conventional level-set formulation under laboratory hot-torsion tests and industrial multi-pass hot rolling conditions. After a limited recalibration of grain boundary mobility and solute drag parameters, both solvers provide comparable predictions of recrystallization kinetics, grain size evolution and final microstructures. The new solver achieves a reduction in computational cost close to two orders of magnitude, while preserving the predictive capabilities of DIGIMU®, thereby enabling more efficient industrial-scale simulations. Simulated predictions will be compared to Ugitech experimental work on lab torsion tests and industrial extrusion processes.

Abstract: The mechanical analysis of sliding frictional contact under small scales is important to improve the understanding about the influence of the contact conditions on the real contact area and, consequently, on the apparent coefficient of friction. This study uses the finite element method to model the contact between an elastoplastic body and a rigid surface with a unidirectional sinusoidal topography, including large sliding. A sensitivity analysis is presented, studying the influence of the initial average pressure, local coefficient of friction and asperity wavelength on the contact conditions. The ratio between total tangential and normal force (apparent friction coefficient) reaches a steady state after a sliding distance of five roughness wavelengths, except for lower values of average initial contact pressure. Increasing the initial average contact pressure leads to an increase of the steady state apparent friction coefficient, particularly for a surface with sharper asperities. This increasing tends to stagnate also with the increase of the local friction coefficient. Withing the cases studied, increasing the initial average contact pressure from 25% to 100% of the material yield stress, leads to an increase of up to 0.07 in the apparent coefficient of friction and of the real-to-apparent contact area ratio up to 30%.